def eval_ckpt():
parser = argparse.ArgumentParser()
parser.add_argument("config_paths", nargs="+", help="path to config.yaml")
parser.add_argument("--weight", help="path to weight to evaluate.pth")
parser.add_argument("--result_dir", help="path to save the result file")
args, left_argv = parser.parse_known_args()
cfg = Config(*args.config_paths, default="cfgs/defaults.yaml")
cfg.argv_update(left_argv)
img_dir = Path(args.result_dir)
img_dir.mkdir(parents=True, exist_ok=True)
trn_transform, val_transform = setup_transforms(cfg)
g_kwargs = cfg.get('g_args', {})
gen = Generator(1, cfg.C, 1, **g_kwargs).cuda()
weight = torch.load(args.weight)
if "generator_ema" in weight:
weight = weight["generator_ema"]
gen.load_state_dict(weight)
test_dset, test_loader = get_test_loader(cfg, val_transform)
for batch in test_loader:
style_imgs = batch["style_imgs"].cuda()
char_imgs = batch["source_imgs"].unsqueeze(1).cuda()
out = gen.gen_from_style_char(style_imgs, char_imgs)
fonts = batch["fonts"]
chars = batch["chars"]
for image, font, char in zip(refine(out), fonts, chars):
(img_dir / font).mkdir(parents=True, exist_ok=True)
path = img_dir / font / f"{char}.png"
save_tensor_to_image(image, path)
def test_dumps_coloring():
dic = yaml.load("""
a: 10
b: 20
""")
cfg = Config(dic, colorize_modified_item=True)
cfg.argv_update(['--a', '20'])
assert cfg.dumps() == "\033[36ma: 20\n\033[0mb: 20"
assert cfg.dumps() == "\x1b[36ma: 20\n\x1b[0mb: 20" # hexa
assert cfg.dumps(modified_color=None) == "a: 20\nb: 20"
def test_none_cli():
dic = yaml.load("""
a: 1
b: 2
c: 3
""")
cfg = Config(dic)
cfg.argv_update(
['--a', 'None', '--b', 'NONE', '--c', 'none', '--d', 'NOne'])
assert cfg['a'] == 'None'
assert cfg['b'] == 'NONE'
assert cfg['c'] == 'none'
assert cfg['d'] == 'NOne'
def test_true_cli():
dic = yaml.load("""
test: True
hmm: true
a:
q: TRUE
w: tRuE
""")
cfg = Config(dic)
cfg.argv_update(
['--test', 'true', '--hmm', 'True', '--a.q', 'trUE', '--a.w', 'TRUE'])
assert cfg['test'] is True
assert cfg['hmm'] is True
assert cfg['a']['q'] == 'trUE'
assert cfg['a']['w'] is True
def test_null_cli():
dic = yaml.load("""
a: 1
b: 2
c: 3
""")
cfg = Config(dic)
cfg.argv_update([
'--a', 'null', '--b', 'Null', '--c', 'NULL', '--d', 'NUll', '--e',
'\'Null\''
])
assert cfg['a'] is None
assert cfg['b'] is None
assert cfg['c'] is None
assert cfg['d'] == 'NUll'
assert cfg['e'] == 'Null'
def test_false_cli():
dic = yaml.load("""
a: false
b: FAlse
c: faLSE
""")
cfg = Config(dic)
cfg.argv_update([
'--a', 'False', '--b', 'false', '--c', 'FALSE', '--d', 'Fals', '--e',
'FAlse'
])
assert cfg['a'] is False
assert cfg['b'] is False
assert cfg['c'] is False
assert cfg['d'] == 'Fals'
assert cfg['e'] == 'FAlse'
def setup_args_and_config():
parser = argparse.ArgumentParser()
parser.add_argument("config_paths", nargs="+", help="path/to/config.yaml")
args, left_argv = parser.parse_known_args()
cfg = Config(*args.config_paths,
default="cfgs/defaults.yaml",
colorize_modified_item=True)
cfg.argv_update(left_argv)
if cfg.use_ddp:
cfg.n_workers = 0
cfg.work_dir = Path(cfg.work_dir)
(cfg.work_dir / "checkpoints").mkdir(parents=True, exist_ok=True)
return args, cfg
def setup_args_and_config():
parser = argparse.ArgumentParser('MaHFG')
parser.add_argument("name")
parser.add_argument("config_paths", nargs="+")
parser.add_argument("--show", action="store_true", default=False)
parser.add_argument("--resume", default=None)
parser.add_argument("--finetune", default=None)
parser.add_argument("--log_lv", default='info')
parser.add_argument("--debug", default=False, action="store_true")
parser.add_argument("--tb-image",
default=False,
action="store_true",
help="Write image log to tensorboard")
parser.add_argument("--deterministic", default=False, action="store_true")
args, left_argv = parser.parse_known_args()
assert not args.name.endswith(".yaml")
cfg = Config(*args.config_paths, colorize_modified_item=True)
cfg.argv_update(left_argv)
if args.debug:
cfg['print_freq'] = 1
cfg['tb_freq'] = 1
cfg['max_iter'] = 10
# cfg['save'] = 'last'
cfg['val_freq'] = 5
cfg['save_freq'] = 10
args.name += "_debug"
args.tb_image = True
args.log_lv = 'debug'
cfg['data_dir'] = Path(cfg['data_dir'])
assert cfg['save_freq'] % cfg['val_freq'] == 0
return args, cfg
def setup_args_and_config():
parser = argparse.ArgumentParser()
parser.add_argument("name")
parser.add_argument("config_paths", nargs="+", help="path/to/config.yaml")
parser.add_argument("--resume", default=None, help="path/to/saved/.pth")
parser.add_argument("--use_unique_name", default=False, action="store_true", help="whether to use name with timestamp")
args, left_argv = parser.parse_known_args()
assert not args.name.endswith(".yaml")
cfg = Config(*args.config_paths, default="cfgs/defaults.yaml",
colorize_modified_item=True)
cfg.argv_update(left_argv)
if cfg.use_ddp:
cfg.n_workers = 0
cfg.work_dir = Path(cfg.work_dir)
cfg.work_dir.mkdir(parents=True, exist_ok=True)
if args.use_unique_name:
timestamp = utils.timestamp()
unique_name = "{}_{}".format(timestamp, args.name)
else:
unique_name = args.name
cfg.unique_name = unique_name
cfg.name = args.name
(cfg.work_dir / "logs").mkdir(parents=True, exist_ok=True)
(cfg.work_dir / "checkpoints" / unique_name).mkdir(parents=True, exist_ok=True)
if cfg.save_freq % cfg.val_freq:
raise ValueError("save_freq has to be multiple of val_freq.")
return args, cfg
def eval_ckpt():
import argparse
from models import generator_dispatch
from sconf import Config
from train import setup_transforms
from datasets import load_json, get_fact_test_loader
logger = Logger.get()
parser = argparse.ArgumentParser()
parser.add_argument("config_paths", nargs="+", help="path to config.yaml")
parser.add_argument("--weight", help="path to weight to evaluate.pth")
parser.add_argument("--img_dir", help="path to save images for evaluation")
parser.add_argument(
"--test_meta",
help=
"path to metafile: contains (font, chars (in unicode)) to generate and reference chars (in unicode)"
)
args, left_argv = parser.parse_known_args()
cfg = Config(*args.config_paths, default="cfgs/defaults.yaml")
cfg.argv_update(left_argv)
content_font = cfg.content_font
n_comps = int(cfg.n_comps)
trn_transform, val_transform = setup_transforms(cfg)
env = load_lmdb(cfg.data_path)
env_get = lambda env, x, y, transform: transform(
read_data_from_lmdb(env, f'{x}_{y}')['img'])
test_meta = load_json(args.test_meta)
dec_dict = load_json(cfg.dec_dict)
g_kwargs = cfg.get('g_args', {})
g_cls = generator_dispatch()
gen = g_cls(1, cfg['C'], 1, **g_kwargs, n_comps=n_comps)
gen.cuda()
weight = torch.load(args.weight)
if "generator_ema" in weight:
weight = weight["generator_ema"]
gen.load_state_dict(weight)
logger.info(f"Resumed checkpoint from {args.weight}")
writer = None
evaluator = Evaluator(env, env_get, logger, writer, cfg["batch_size"],
val_transform, content_font)
img_dir = Path(args.img_dir)
ref_unis = test_meta["ref_unis"]
gen_unis = test_meta["gen_unis"]
gen_fonts = test_meta["gen_fonts"]
target_dict = {f: gen_unis for f in gen_fonts}
loader = get_fact_test_loader(env,
env_get,
target_dict,
ref_unis,
cfg,
None,
dec_dict,
val_transform,
ret_targets=False,
num_workers=cfg.n_workers,
shuffle=False)[1]
logger.info("Save CV results to {} ...".format(img_dir))
evaluator.save_each_imgs(gen,
loader,
save_dir=img_dir,
phase="fact",
reduction='mean')
def eval_ckpt():
from train import (setup_language_dependent, setup_data,
setup_cv_dset_loader, get_dset_loader)
logger = Logger.get()
parser = argparse.ArgumentParser('MaHFG-eval')
parser.add_argument(
"name",
help=
"name is used for directory name of the user-study generation results")
parser.add_argument("resume")
parser.add_argument("img_dir")
parser.add_argument("config_paths", nargs="+")
parser.add_argument("--show", action="store_true", default=False)
parser.add_argument(
"--mode",
default="eval",
help="eval (default) / cv-save / user-study / user-study-save. "
"`eval` generates comparable grid and computes pixel-level CV scores. "
"`cv-save` generates and saves all target characters in CV. "
"`user-study` generates comparable grid for the ramdomly sampled target characters. "
"`user-study-save` generates and saves all target characters in user-study."
)
parser.add_argument("--deterministic", default=False, action="store_true")
parser.add_argument("--debug", default=False, action="store_true")
args, left_argv = parser.parse_known_args()
cfg = Config(*args.config_paths)
cfg.argv_update(left_argv)
torch.backends.cudnn.benchmark = True
cfg['data_dir'] = Path(cfg['data_dir'])
if args.show:
exit()
# seed
np.random.seed(cfg['seed'])
torch.manual_seed(cfg['seed'])
random.seed(cfg['seed'])
if args.deterministic:
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
cfg['n_workers'] = 0
logger.info("#" * 80)
logger.info("# Deterministic option is activated !")
logger.info(
"# Deterministic evaluator only ensure the deterministic cross-validation"
)
logger.info("#" * 80)
else:
torch.backends.cudnn.benchmark = True
if args.mode.startswith('mix'):
assert cfg['g_args']['style_enc']['use'], \
"Style mixing is only available with style encoder model"
#####################################
# Dataset
####################################
# setup language dependent values
content_font, n_comp_types, n_comps = setup_language_dependent(cfg)
# setup transform
transform = transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize([0.5], [0.5])])
# setup data
hdf5_data, meta = setup_data(cfg, transform)
# setup dataset
trn_dset, loader = get_dset_loader(hdf5_data,
meta['train']['fonts'],
meta['train']['chars'],
transform,
True,
cfg,
content_font=content_font)
val_loaders = setup_cv_dset_loader(hdf5_data, meta, transform,
n_comp_types, content_font, cfg)
#####################################
# Model
####################################
# setup generator only
g_kwargs = cfg.get('g_args', {})
gen = MACore(1,
cfg['C'],
1,
**g_kwargs,
n_comps=n_comps,
n_comp_types=n_comp_types,
language=cfg['language'])
gen.cuda()
ckpt = torch.load(args.resume)
logger.info("Use EMA generator as default")
gen.load_state_dict(ckpt['generator_ema'])
step = ckpt['epoch']
loss = ckpt['loss']
logger.info("Resumed checkpoint from {} (Step {}, Loss {:7.3f})".format(
args.resume, step, loss))
writer = utils.DiskWriter(args.img_dir, 0.6)
evaluator = Evaluator(hdf5_data,
trn_dset.avails,
logger,
writer,
cfg['batch_size'],
content_font=content_font,
transform=transform,
language=cfg['language'],
val_loaders=val_loaders,
meta=meta)
evaluator.n_cv_batches = -1
logger.info("Update n_cv_batches = -1 to evaluate about full data")
if args.debug:
evaluator.n_cv_batches = 10
logger.info("!!! DEBUG MODE: n_cv_batches = 10 !!!")
if args.mode == 'eval':
logger.info("Start validation ...")
dic = evaluator.validation(gen, step)
logger.info("Validation is done. Result images are saved to {}".format(
args.img_dir))
elif args.mode.startswith('user-study'):
meta = json.load(open('meta/kor-unrefined.json'))
target_chars = meta['target_chars']
style_chars = meta['style_chars']
fonts = meta['fonts']
if args.mode == 'user-study':
sampled_target_chars = uniform_sample(target_chars, 20)
logger.info("Start generation kor-unrefined ...")
logger.info("Sampled chars = {}".format(sampled_target_chars))
evaluator.handwritten_validation_2stage(gen,
step,
fonts,
style_chars,
sampled_target_chars,
comparable=True,
tag='userstudy-{}'.format(
args.name))
elif args.mode == 'user-study-save':
logger.info("Start generation & saving kor-unrefined ...")
save_dir = Path(args.img_dir) / "{}-{}".format(args.name, step)
evaluator.handwritten_validation_2stage(gen,
step,
fonts,
style_chars,
target_chars,
comparable=True,
save_dir=save_dir)
logger.info("Validation is done. Result images are saved to {}".format(
args.img_dir))
elif args.mode == 'cv-save':
save_dir = Path(args.img_dir) / "cv_images_{}".format(step)
logger.info("Save CV results to {} ...".format(save_dir))
utils.rm(save_dir)
for tag, loader in val_loaders.items():
l1, ssim, msssim = evaluator.cross_validation(
gen,
step,
loader,
tag,
n_batches=evaluator.n_cv_batches,
save_dir=(save_dir / tag))
else:
raise ValueError(args.mode)
